Abstract
BACKGROUND: Machine learning (ML) is combined with noninvasive parameters from coronary computed tomography angiography (CTA) to construct predictive models to identify culprit lesions that may lead to acute coronary syndrome (ACS). METHODS: We retrospectively analyzed 132 patients with ACS at the Fourth Affiliated Hospital of Harbin Medical University who had coronary CTA between 3 months and 3 years before the ACS event, with a total of 240 lesions. Lesions from 2020 (n=154) were included in the training set, and lesions from 2021 (n=86) were included in the test set for internal validation. We evaluated the role of plaque characteristics, hemodynamic parameters and pericoronary adipose tissue (PCAT) attenuation from CTA in identifying culprit ACS lesions. In the training set, logistic regression was used to screen CTA-derived parameters with P values <0.05 for the model construction. Logistic regression, random forest, Bayesian and K-nearest neighbor algorithms were used to build classification models, and their performance was assessed using the test set. The following models were established to evaluate the effectiveness of different combinations of models to identify culprit lesions: Model 1 was established for plaque characteristics; Model 2 was established for hemodynamic parameters; Model 3 was established for PCAT attenuation; Model 4 was established for plaque characteristics and hemodynamic parameters; and Model 5 was established for plaque characteristics, hemodynamic parameters and PCAT attenuation. RESULTS: A total of ten high-risk factors were screened for the ML model construction, and the ML model based on the logistic regression algorithm had the best performance among the four algorithms (accuracy =0.721; sensitivity =0.892; specificity =0.592; positive prediction =0.623; and negative prediction =0.879). In this model, the minimum lumen area, positive remodeling and lesion-specific fat attenuation index (FAI) were the risk factors significantly associated with the culprit lesion. Analysis of the effect of different combinations of models to identify culprit lesions showed that Model 5 had the best predictive effect (AUC =0.819 and 95% CI: 0.722-0.916). CONCLUSIONS: ACS can be predicted using ML based on CTA parameters. Compared to other models, the model combining plaque characteristics, hemodynamic parameters and PCAT attenuation performed best in predicting the culprit lesion.